Bayesian Optimisation Lite

Read this before using it: this header is NOT a Gaussian process and does not claim to be. Real Bayesian optimisation fits a Gaussian process posterior over the objective function, which needs inverting a covariance matrix that grows with every observation - genuinely hard to implement correctly and keep numerically stable in plain MQL4 with no linear-algebra library. Rebuilding that honestly was judged out of scope for what this header can responsibly deliver, so a documented simplification is used instead. The simplification: a Nadaraya-Watson kernel-weighted average of nearby observed objective values stands in for the GP posterior MEAN, and a distance-to-nearest-observation proxy (scaled by the observed objective spread) stands in for the GP posterior UNCERTAINTY. Both use the same RBF kernel a real GP would use as its covariance function, so the surrogate still concentrates predicted uncertainty in genuinely under-explored regions - but the numbers are a heuristic, not a calibrated statistical posterior. Ranking candidate parameter values by Expected Improvement over this cheap surrogate still produces the same qualitative behaviour real Bayesian optimisation is used for: early evaluations spread out to explore, later ones concentrate near the best region found so far, converging on a good parameter value in tens of evaluations rather than an exhaustive scan of hundreds or thousands of strategy-tester runs. This is a single-parameter (1D) optimiser. A multi-dimensional version would need the same kernel machinery extended to a distance metric over several parameters at once, which was left out to keep the surrogate's behaviour easy to reason about and verify - tuning one parameter at a time against a fixed baseline for the others is also, in practice, how most traders actually explore an EA's parameter space. Notes and limitations - Not a Gaussian process. The uncertainty figure it reports is a distance- based heuristic, not a calibrated confidence interval - do not use it to make statistical claims, only to guide which region to try next. - Single parameter only. Optimising several parameters together needs either running this once per parameter with the others held fixed, or a genuinely different multi-dimensional implementation this header does not provide. - The surrogate mean can be a poor predictor far from any observed point (it falls back to the plain observed average there) - the algorithm compensates by ALSO ranking those points as highly uncertain, which is what keeps early search spread out rather than stuck near the first points evaluated, but it is a heuristic compensation, not a guarantee. - Best suited to smooth, single-peaked or few-peaked objective landscapes. A genuinely noisy or highly multi-modal objective can mislead the kernel- weighted mean the same way it would mislead a real GP with too coarse a length scale - try a smaller InpLengthScale if the search seems to settle on the wrong region. - This header never calls the strategy tester itself - EvaluateMyStrategy in the usage example above is the caller's own responsibility to wire in.

Script · MT5 · v1.0 · Free

Tags: MT5 Script · MQL5 Source Code

MQL4 & MQL5 scripts · PolyMaria